Replace qwen3-asr custom helper with optimum-intel OVModelForSpeechSeq2Seq

[openvino-dev-samples]

What does this PR do?

Replaces the custom qwen_3_asr_helper.py (1442 lines of manual model conversion and inference code) with the standard optimum-intel OVModelForSpeechSeq2Seq API, following the pattern established in huggingface/optimum-intel#1677.

Changes

• Delete qwen_3_asr_helper.py — removes custom OpenVINO conversion code, manual GenerationMixin wrapper, and custom inference pipeline
• Rewrite qwen3-asr.ipynb — use \OVModelForSpeechSeq2Seq.from_pretrained(export=True)\ for export and \OVModelForSpeechSeq2Seq\ for inference
• Simplify gradio_helper.py — use standard processor + model.generate() pattern instead of custom transcribe API
• Update README.md — reflect new optimum-intel based workflow

Key benefits

• ~1600 lines deleted of fragile custom code (manual model tracing, stateful model patching, custom KV-cache management, etc.)
• Standard API — uses the same \OVModelForSpeechSeq2Seq\ pattern as distil-whisper-asr and other speech notebooks
• Easier maintenance — model export and inference handled by optimum-intel library
• No Qwen3-ASR repo clone needed — only \pip install qwen-asr\ required to register the model

Dependencies

• Requires optimum-intel with Qwen3-ASR support (merged in huggingface/optimum-intel#1677). Currently uses \git+https://github.com/huggingface/optimum-intel.git\ until the next release.
• \ ransformers>=4.53, \openvino>=2025.4.0, \qwen-asr\

Testing

Tested end-to-end in a fresh virtual environment:

1. Export Qwen3-ASR-0.6B to OpenVINO format ✅
2. Load model and processor ✅
3. Run inference on English sample audio ✅

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB